Convolutional gated MLP and attention improve end-to-end spoken language understanding

The task of spoken language understanding(SLU) consists mainly of intention detection and slot filling. Most studies model the two separately, with intention detection as a categorization task and slot filling as a sequence labeling task, which are interrelated. End-to-end(E2E) approaches have recently received much attention due to their structural simplicity and effectiveness. Still, the need for speech datasets is one of the main reasons for their limited development. Second, E2E SLU is plagued by long-tail words. To mitigate the above problems, the use of Hubert as a speech encoder combined with the Convolutional Gated MLP(CgMLP) module is proposed to improve the feature learning capability of the model. The intention detection and slot-filling tasks are jointly modeled E2E. At the same time, the long-tailed word problem is mitigated using the attention mechanism at the decoding, an improvement of 4.77\% and 1.87\% over the baseline model on the two tasks.